A Friend — Distribution

references__essence-extraction-framework

9 min read

Synced from references/essence-extraction-framework.md in content-extraction skill on 2026-05-18. Edit upstream in the skill; this file is overwritten on next sync.

Essence-Extraction Framework — Schärfung / Trim / Essenz

The operational framework the virality sub-agent applies to Daniel’s source. Owned by virality/SKILL.md. The German triple — Schärfung (sharpening), Trim, Essenz-Extraktion (essence-extraction) — is Daniel’s framing (verbatim 2026-05-03). This file makes those three actions concrete.

The premise: a raw 2-3 minute monologue contains 30-60s of actually-good content. The sub-agent’s job is to find that 30-60s, cut everything else, and surface what specifically pulls. Length is not a virtue. A 50s clip with substance pulls harder than a 115s clip with substance + fluff.

Three operations, applied in order:

  1. Schärfung — find the strongest punches in source; pull them to clip-second 0
  2. Trim — drop weak material adjacent to the punches; cut redundancy WITHIN strong segments
  3. Essenz — distill the kernel of the clip’s argument; verify the I-believe-that landed

This file defines the heuristics each operation applies. The sub-agent runs each operation against the draft cut-plan arcs (post-silence-removal, pre-cut) and proposes specific changes in virality-pass.md.

Operation 1: Schärfung — find the strongest punch, lead with it

Premise: A 2-3min Daniel monologue typically has 1-3 Tier-1 self-supplying scroll-stoppers. They’re rarely at the start of source. They’re usually in the LAST third (Daniel’s cadence — he warms up, then lands the kernel).

Where to look (in order of yield):

  1. Last 30s of source. Daniel’s strongest reframes statistically land here. His “warm-up → land” arc means the last beat IS often the kernel. Already encoded in parent SKILL.md Rule A “Tail-of-source as opener” — sub-agent enforces explicitly.

  2. Mid-source vulnerability beats. When Daniel says “I almost…” / “I was wrong about…” / “this can’t be it” — those are usually unfiltered. Specificity earned through emotion.

  3. Tool-naming declaratives. “Pulled up all my agents” / “I’m using Claude + Figma + the skill” → Pillar 7 punches. Land hard because they’re concrete + named.

  4. Two-beat cadence in source. Brand’s signature rhythm. Daniel’s: “Tools end conversations. Companions hold them.” These are auto-Tier-1 — pull forward whenever found.

  5. Categorical reframes. “We are not merely building toys” / “this constant overwhelm of having multiple windows of agents” — reframes existing assumptions. Self-supplying.

Where NOT to look:

  • Source intro. Series-anchor language (“welcome to episode N” / “today I want to talk about”) almost never qualifies. Functional, not punchy.
  • Restart-pause openers. “Yeah, so…” / “Okay, basically…” → filler.
  • Generic vulnerability. “Building a startup is hard” without specificity → fails So-What test.

Schärfung outputs (per draft clip):

### Schärfung
- Current opener: "<verbatim>" (segN, source HH:MM:SS)
- **Proposed opener:** "<verbatim from elsewhere in source>" (segM, source HH:MM:SS)
- Why this lands harder: <name the technique — categorical reframe / vulnerability / tool-naming / two-beat / tail-distillation>
- Source-tail searched: ✓ (sub-agent MUST verify last 30s explicitly)

If the proposed opener is literally the original opener — that’s a valid output. But the sub-agent must explicitly state the tail was searched and rejected (don’t skip the work).

Operation 2: Trim — kill weak material, internal + adjacent

Once the punch is identified, every adjacent second must earn its place. Three sub-operations:

Trim 2a: Adjacent-weakness elimination

A weak segment between two strong ones gets cut. The viewer’s retention is determined by the WEAKEST adjacent beat, not the average. A clip with [punch — filler — punch] reads worse than [punch — punch] at half the runtime.

What counts as “weak adjacent”:

  • Restating the previous beat in different words
  • Conjunction-led continuations (“So that’s why…”, “And that’s how…”) without new substance
  • Repetition of theme without new angle
  • Fluffy reflection that doesn’t add specificity (“I think about this a lot”)
  • Topic drift that disrupts the arc

Heuristic: for each segment N, ask “does this segment introduce a NEW substance beat? (named thing / specific number / vivid image / pivot / cliffhanger)” — if no, propose drop.

Trim 2b: Internal redundancy (Rule K extension)

A single segment can have internal redundancy. Daniel’s natural cadence sometimes restates inside a sentence — “I got two more days and I said to myself, two more days, then it’s done.” The repetition + connective tissue (“and I said to myself”) can usually be cut to land tighter.

When to apply:

  • A phrase repeats inside a 30s window (often natural-speech redundancy).
  • Connective tissue (“and I said to myself”, “you know”, “so basically”) sits between two punches.
  • A long sentence has a tight kernel buried in qualifying clauses.

When to NOT apply:

  • The repetition IS the rhetoric (deliberate restatement for emphasis — Daniel’s “two more days. Two more days.” cadence).
  • Connective tissue carries voice (cutting kills it).
  • Sub-cut would break audio flow into a glitch.

Path: word-level data tells you when each word starts/ends. Find redundant span in/out points, encode as sub-segment cut. Validate audibly.

Trim 2b outputs:

### Trim (word-level cuts inside quotes)
- segN quote: "<original>"
- **Drop:** "<phrase>" (word-level: X.XXs–Y.YYs)
- Audio result: tighter [recursive-shrink reveal / punchline landing / cadence]

Trim 2c: Weak-tail elimination

Outro/CTA segments are almost always trimmable. Daniel’s “so let’s get on this journey…” closes have value as series anchors — but for shipped short-form they’re sells-y and break the loopability rule.

Default action: propose dropping outro segments unless they contain a promise-close or two-beat cadence. The clip should END on the climax beat, not the wave-goodbye.

Operation 3: Essenz — what is this clip’s ONE sentence?

Every clip has exactly one I-believe-that line that completes its argument. If the clip can’t be reduced to one sentence Daniel would own, the clip isn’t ready.

The kernel test:

“If this monologue had to fit in one sentence Daniel would still own, which sentence would it be?”

That sentence is the Essenz. It’s:

  • A verbatim line from source (NOT synthesized)
  • Specific enough that someone could quote it back to him and he’d recognize it
  • Actually about ONE thing (not a list)
  • Stake-bearing (passes So-What test)

The kernel comes BEFORE the cut, not after. The sub-agent identifies the Essenz first, then asks: “what segments support this Essenz? what segments distract from it?”

Segments that don’t support the Essenz get cut. Even if they’re well-said. Even if Daniel meant them. They belong in a DIFFERENT clip with a DIFFERENT Essenz.

This is how a 115s master becomes a 50s clip: most of the master’s segments don’t all support the same Essenz. They support multiple Essenz lines that need separate clips.

Essenz outputs:

### Essenz (one-sentence kernel)
> "<the I-believe-that line that completes the clip's argument>"
 
**Source location:** segN, HH:MM:SS
**Pillar fit:** <pillar from _brand/pillars.md>
**Schwartzberg pass:**
- I-believe-that: ✓ "<expanded sentence>"
- So-what: <Yes — who pushes back / No — too obvious>
- Why-test: <weighted vs filler analysis>
 
**Segments supporting this Essenz:** seg2, seg4, seg7
**Segments NOT supporting this Essenz (propose drop):** seg1, seg3, seg5, seg6

If a clip’s source contains TWO viable Essenz lines, the sub-agent proposes splitting into TWO clips. Don’t try to ship a clip that argues two things.

The Compression Math

How much should a clip get cut from raw source? Heuristic:

Source durationTypical shipped durationCompression ratio
30s raw20-25s80% kept
60s raw30-40s50-65% kept
90s raw35-50s40-55% kept
120s raw40-60s35-50% kept
180s raw45-70s25-40% kept

These ratios are rough — substance density varies. But they’re the bias: a 3-minute monologue should yield a sub-1-minute clip in the typical case.

If silence-removal alone (Phase 1) produces a 1.5min cut from a 3min source, the sub-agent’s job is to recommend cutting it FURTHER to 50-60s. The path is: identify Essenz → keep supporting segments → drop the rest, even if the rest is nice.

Anti-pattern: treating Phase 1’s silence-removed output as the final cut. That’s the equivalent of “I removed the dead-air, ship it.” Dead-air removal is necessary but not sufficient. The substance-density check (Essenz alignment) is what shortens the clip from 1.5min to 50s.

The Schärfung-Trim-Essenz integration loop

Per draft clip, sub-agent runs:

  1. Schärfung pass → propose hook reorder
  2. Essenz pass → identify the kernel
  3. Trim pass → propose adjacent + internal cuts that don’t support the Essenz
  4. Recompute target duration → cut should be 30-60s for short-form publish, ≤30s if Essenz is tight, ≤90s only if Essenz justifies depth
  5. Re-run attention-preservation scorecard (references/attention-preservation.md) → verify second-hook lands at 12-18s, pacing-shifts present, loopability achievable

If after all passes the proposed clip still exceeds 60s OR has a flat retention curve, the sub-agent’s verdict is REVIEW or REJECT — Daniel decides. Don’t ship a clip the sub-agent itself flagged as marginal.

What the framework does NOT do

  • Does NOT synthesize new lines. Every proposed opener / kept segment / trim recommendation is verbatim from transcript.txt.
  • Does NOT auto-merge into cut-plan.json. Outputs virality-pass.md for Daniel review.
  • Does NOT enforce hard runtime caps. Recommends a target; Daniel decides.
  • Does NOT preserve Daniel’s intended order. Reorder is on the table — anything verbatim can move.
  • Does NOT score against general short-form virality. Scores against A Friend register (anti-slick — see references/aesthetic-register-anti-slick.md).

Brand non-negotiables (hard gates)

The sub-agent’s output MUST pass these checks before being surfaced to Daniel:

  • Every proposed opener / kept segment / trim is verbatim substring of transcript.txt (whitespace-tolerant).
  • No banned vocab in any sub-agent surface text (titles, framing, commentary). Banned: revolutionize, seamless, AI-powered, game-changer, leverage, exclamation points anywhere.
  • Two-beat cadence preserved in any proposed re-arc.
  • Code-switching preserved (German/English mix).
  • Tool-naming check (Pillar 7 only) — surfaced in virality-pass.md.
  • Trim proposals ≥0.4s tagged “VALIDATE BY EAR” — Daniel listens before merge.

Failures auto-inject BRAND-NONNEGOTIABLE-VIOLATION tag inside virality-pass.md rather than silent emit.

Sources & canon basis

  • references/shortform-virality.md — parent rubric (Tier-1/2/3, hook archetypes, Schwartzberg).
  • references/attention-preservation.md — temporal mechanics (second-hook window, loopability, retention floors).
  • _brand/voice-preservation.md — verbatim contract, banned vocab, two-beat, code-switching, “feels like me” gate.
  • _brand/pillars.md — 7 pillars, anti-pattern detection.
  • Joel Schwartzberg Get to the Point — I-believe-that / So-what / Why-test framework.
  • William Strunk Jr. & E. B. White The Elements of Style — omission discipline (“Omit needless words”).
  • Nathan Barry “compression as care” essay — substance-density per second as the right metric.
  • A Friend campaign-state.md — what’s been shipped, which Essenz lines exist already, anti-duplication.

The framework cites THIS file when documenting a recommendation. Not generic content-creator advice.

Graph View